Repeated decisions, correctly made.

The gap

The current wave of AI has allowed us to treat natural language as code and gain insights from free-form text. That is a real advance. It does not tell a firm what action to take.

A price, a replenishment, a route, a credit line, or the next step in a funnel is not a classification problem, nor a next-token prediction problem. Actions change the state of the world. Correct decisions need systems that can both predict the outcome of actions and plan more than one step ahead. The objective is return over the trajectory, not accuracy on a held-out sentence.

Firms already record these trajectories. The data is operational, not linguistic. It is collected as a byproduct of software and rarely treated as something that can be modeled and planned against. Current methods are weak on this class of sequential optimization, yet making these decisions correctly is what unlocks business value.

Approach

The big challenge that we don’t see in our field yet is the ability to learn a model and then plan with a model.

That is the technical program. Learn a model of the decision process from historical actions and outcomes. Plan against that model under the constraints the business already has. Compare the resulting policy with the current rule, limit how far it may deviate, and evaluate it on return — not on a proxy loss.

This work sits with operations research and model-based reinforcement learning. Language models can sit at the edge of the system. They are not the decision procedure.

Background

We bring decades of experience putting machine learning into production systems, and direction-setting research at NeurIPS, ICML, AISTATS, UAI, KDD, and RecSys, and in JMLR and TMLR.

Focus

We take problems where a decision is already being made often, the revenue attached to it is large, and integration is a data path rather than a new sensor network.

Problem classes

Funnel optimization

Pricing

Logistics

Industries

Retail

Travel

Financial services